Will Depue reports early signs of GPT-6 Astra auto-interp progress on Qwen3.5-0.8B

OpenAI researcher Will Depue says there are early signs of life from Astra-next's auto-interpretability work, though he describes the latest results as not dramatic.

He reports the work on Qwen3.5-0.8B weights, which he calls Weight-Space Programming results. Depue says the next step is upstream reinforcement learning tasks to make Astra-next much better at auto-interp work.